[
https://issues.apache.org/jira/browse/IMPALA-5050?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16618050#comment-16618050
]
ASF subversion and git services commented on IMPALA-5050:
---------------------------------------------------------
Commit 2ee8caeb3053dfa2c434c680ffb2ac756627ee38 in impala's branch
refs/heads/master from [~csringhofer]
[ https://git-wip-us.apache.org/repos/asf?p=impala.git;h=2ee8cae ]
IMPALA-7521: Speed up sub-second unix time->TimestampValue conversions
Impala used to convert from sub-second unix time to TimestampValue
(which is split to date_ and time_ similarly to
boost::posix_time::ptime) by first splitting the input into seconds
and sub-seconds part, converting the seconds part wit
boost::posix_time::from_time_t(), and then adding the sub-seconds
part to this timestamp.
Different tricks are used to speed up different functions:
- UTC functions that expect a single integer as input can
split it into date_ and time_ directly.
- Non-UTC functions need seconds for timezone conversion,
because CCTZ expects time points as seconds. These
were optimized by adding the subsecond part to time_
instead of adding it to a ptime. This can be done safely
because the sub-second part is between [0, 1 sec), so
it cannot overflow into a different day or timezone.
Benchmarks show 2x - 6x speedup for the measured functions.
The main motivation is IMPALA-5050: "Add support to read
TIMESTAMP_MILLIS and TIMESTAMP_MICROS to the parquet
scanner" - reading these types will run
micro/milli->TimestampValue conversion for every row.
Other changes:
- TimestampValue::UtcFromUnixTimeMillis was added - currently this
is only used in tests but it will be useful for IMPALA-5050
- Some functions were moved from .h to .inline.h.
- FromUnixTimeMicros was changed to do the utc->local conversion
depending on flag use_local_tz_for_unix_timestamp_conversions
to be consistent with other similar functions. This function was
only used in tests until now but it will be useful for IMPALA-5050.
- When a result mismatch is detected in
convert-timestamp-benchmark.cc it now prints non-equal values.
- Benchmarks were added for micro + nano conversions.
Note that only single threaded benchmarks were added because I
do not expect any difference in the multi threaded case.
- DCHECKs were added to TimeStampValue::Validate to
ensure that time_ is between [0, 24 hour).
Testing:
- timestamp-test.cc was extended to give better coverage
for sub-second conversions. Edge cases were already
covered pretty well.
Change-Id: I572b5876b979ddae58165bd40d5b008ce9d7a4aa
Reviewed-on: http://gerrit.cloudera.org:8080/11183
Reviewed-by: Impala Public Jenkins <impala-public-jenk...@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenk...@cloudera.com>
> Add support to read TIMESTAMP_MILLIS and TIMESTAMP_MICROS to the parquet
> scanner
> --------------------------------------------------------------------------------
>
> Key: IMPALA-5050
> URL: https://issues.apache.org/jira/browse/IMPALA-5050
> Project: IMPALA
> Issue Type: New Feature
> Components: Backend
> Affects Versions: Impala 2.9.0
> Reporter: Lars Volker
> Assignee: Csaba Ringhofer
> Priority: Major
>
> This requires updating {{parquet.thrift}} to a version that includes the
> {{TIMESTAMP_MICROS}} logical type.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
